AutoQuorum : Workload-aware Optimization for Replicated State Machines

Sammanfattning

Distributed systems are widely used to design low-latency, scalable, and available services by distributing tasks across multiple machines or nodes. These systems achieve high availability through fault tolerance, allowing them to continue operating even when individual nodes fail, and minimize latency by deploying nodes geographically, enabling client requests to be processed closer to their source. However, when services require consistent state across nodes, the coordination necessary to ensure consistency is at odds with the goal of low latency. The standard approach to achieving consistency is through Replicated State Machines, where operations on shared state must be coordinated among majorities of nodes, known as quorums. This coordination significantly diminishes the latency gains achieved by placing nodes near client requests. This work identifies key configuration parameters that influence quorum coordination in Replicated State Machines and develops techniques to safely and efficiently reconfigure them to optimize latency while maintaining strong consistency. Additionally, a distributed and strongly consistent quorum read algorithm tailored to this dynamic configuration setting is presented. These contributions are implemented in AutoQuorum, a runtime optimizer for leader-based Replicated State Machine protocols, which continuously reconfigures the system to optimize latency in response to changing workloads. AutoQuorum was deployed across geographically distributed Google Cloud Platform instances, and its latency optimization was evaluated across various workloads. The results demonstrate that AutoQuorum can significantly improve latency, reducing it by more than 85% in certain scenarios.